Corrections

Everything we have got wrong on this site, what it said before, and why it changed. Newest first.

Version 1.9 · Published 5 September 2026 · Last updated 7 September 2026 · Covers: every page on this site, since the first one went up on 3 September 2026

What we count as a correction

A correction is a published statement that was wrong, unsourced, or described our own system inaccurately. It gets fixed in place, and it gets a dated entry here saying what it used to say. We do not edit a published claim silently and we do not delete the old wording — a correction you cannot check is not a correction.

A version change is different, and it is not the same thing as an error. If we change a scoring rule, an exclusion rule, the domain model or the norming rung, the meaning of a result changes, so the instrument gets a new version number and old results stay attached to the version they were produced under. The instrument register carries those versions.

Something can be both. Most of what is below is neither — it is us having written a sentence we could not support, and then checking.

Why this page is likely to look bad, and stays anyway

A corrections page grows in one direction. Ours will get longer, and every entry on it is a thing we published and should not have. The alternative is not a site with no errors — it is a site with the same errors and no page listing them.

If you find something wrong here, tell us and it will appear on this page with a date.

20 September 2026 (sixth entry) — three edits to make the privacy notice easier to read

A revision, not a correction. Nothing on the page was wrong; it was harder to read than it needed to be.

• The notice stopped citing its own version numbers. Three sentences said things like “version 1.7 promised…”. A reader needs the promise, not the version that made it, so they now simply state it. The version stamp and this page remain the record of what changed and when.
• Two internal counter names were removed and the point was kept. The notice used to print the literal keys of two old tallies. What matters is not the key names but why they are still there: two anonymous increments of unknown origin, which we cannot prove came from our own testing, and which we therefore left alone rather than deleting to tidy the story. Two others were deleted on 7 September 2026 because we could show they were ours. That distinction is the whole point and it stays.
• The opening line no longer outruns the rest of the page. It said “We collect nothing about you directly”, then a few lines later described analytics and the IP address any server necessarily receives. Both were true and the first read stronger than the rest supported. It now says we do not collect personal information from you directly, and names the two things that happen anyway before listing them.

All three came from an outside reader who went through the live page and said, in effect, that the transparency was no longer the problem — the readability was.

20 September 2026 (fifth entry) — “no advertising” was a promise we had not actually made

A revision, not a correction — but a revision that makes the page less flattering, which is the direction worth being suspicious of when it does not happen.

• What it said. “No advertising and no tracking”, in the present tense, with nothing said about whether that was permanent.
• Why that was a problem. It was true, and it read like a commitment. BrainScoreDaily is a business and advertising is one of the ways it might eventually pay for itself — so a reader could have taken a description of today as a promise about always, and been quietly let down later.
• What it says now. No advertising today, and a section setting out what was ruled out in advance: never on a test and never on a result, only on the pages of text if anywhere, never priced on anything about you, and this notice changes before the first advertisement rather than after.
• The rule about tests and results is not new and is not softening. It was written down before this site had any traffic to sell, specifically so it could not be traded away once there was.

The version-by-version history that used to sit in the footer of the privacy notice has also been removed. It was a development log in the footer of a legal page. This page is where that record belongs, and the notice links here from the top.

20 September 2026 (fourth entry) — the address on this site for privacy questions had never worked

This is the worst defect this site has had. Not a sentence that went out of date — a mechanism that never worked at all, on the page whose whole purpose is to be relied on.

• What was published. privacy@brainscoredaily.com, named on the privacy notice as the way to ask what is held, to have it corrected or erased, and to raise a concern.
• What actually happened to mail sent there. Nothing. brainscoredaily.com had no MX record — no mail exchanger, no route in. Messages were discarded at the internet's own plumbing, and because there was no server to reject them, many senders would not even have received a bounce. Someone exercising a data right would have written into silence and reasonably concluded they were being ignored.
• How long. For as long as the address has been published. We cannot say how many people wrote, because there is no record of a message that was never accepted.
• Why it happened. Cloudflare Email Routing was switched on for the domain and its DNS records were never written — the dashboard had been reading “DNS records: Not configured” and “Routing rules: 0”. A related note in our own records said no mail record had been added, but that note was about sending mail, and nobody asked what it meant for receiving.
• How it was found. Two outside reviewers read this page closely and neither could have caught it; a rendered page cannot tell you whether its own contact address resolves. It surfaced only because, before publishing a block naming that address as the way to reach us, we checked that it worked.
• Fixed the same day. The five missing records were added, a rule now sends privacy@ to a monitored inbox, and a message was sent to the address and arrived — the dashboard saying Active is a claim, a delivered message is evidence.

If you wrote to us before 20 September 2026 and heard nothing back, that is why. We are sorry, and we would be glad to hear from you again.

The lesson is narrower and more uncomfortable than the others on this page. Every previous entry here was a claim that drifted out of date. This one was never true, and it survived because a published contact address looks identical whether it works or not. A promise that depends on a mechanism has to be tested through the mechanism, not proofread.

20 September 2026 (third entry) — the privacy notice was rewritten to be read, not audited

This one is a revision, not a correction. Nothing on the page was wrong. It is here because the page used to carry its own change history in boxes at the top, those boxes are gone, and this is where that record now lives.

• What prompted it. An outside reader went through the live page and said, in effect, that it had stopped being a privacy notice and become a development log. They were right and the number is unflattering: nineteen per cent of the page ran before the words “What we collect”, and all of it was the story of how the page came to be written.
• What moved. Three dated boxes describing changes made on 19 and 20 September. Every one of them duplicated an entry already on this page, so nothing was lost by removing them.
• What was added. An In short summary at the top, so the answer to “what do you collect” arrives in about thirty seconds. A plain statement of how long anything is kept. And a hard heading — Planned — not active yet — around everything that does not exist, because the notice was describing a live site and a future account in one voice.
• One published commitment got narrower. A saved result was going to be able to carry “any note you typed about the conditions”. That free-text box is gone before it was ever built, replaced by a short fixed list — tired, rushed, interrupted, unwell, none of these. A box invites a name, a diagnosis or a circumstance, and the safest way to promise we will not hold those is to build somewhere they cannot be typed.

The reason this entry exists at all is that the boxes came off the notice. A page that carries its own history is harder to read; a page whose history is nowhere is worse. So the history moved here and the notice now links to it from the top.

20 September 2026 (second entry) — we changed how the analytics is installed and left the page describing the old way

The same defect as the analytics entry the day before, with one difference that makes it worse: this time we knew.

• What it said. That the analytics counts were “sent to an address on this domain rather than to a third party”.
• Why it was true when written. Cloudflare can insert its counting script automatically, and when it does, the counts are sent back to an address on this site’s own domain. That is a real privacy advantage and it is why we chose that method.
• What changed. On 19 September we switched to installing the script ourselves, so that we could keep it off the seven tests and off your Fingerprint and have an automated check enforce that. A self-installed script reports to Cloudflare instead. We gained a guarantee and gave up the first-party address.
• What we got wrong. Not the decision — the trade was written down the same day and we would make it again. What we got wrong is that we wrote the trade into our internal record and left this page describing the arrangement we had just stopped using.
• How it was found. An outside reader reviewed this page and said the vendor claims should be checked against the actual configuration. They did not know this one was wrong; they knew it was the kind of claim that goes wrong.
• What changed on the page. The sentence now says the counts go to Cloudflare, because they do.

The lesson we are taking is narrow and specific. A decision that records a trade-off is not the same as a page that reflects it. When a change has a cost, the record of the cost has to reach every surface that made the old promise — and on this site the surfaces are listed, so there is no excuse for not walking them.

20 September 2026 — a list that said “in full” was not full

Smaller than the last one, and found the same way: by checking a promise against what keeping it would actually require.

• What it said. The privacy notice lists what an account would hold under the heading “What we would hold, in full”: your email address, a random account number, the date you created it, and the results you chose to save.
• What was missing. The date you last signed in. The same notice promises that if you do not sign in for eighteen months we will email you once, wait thirty days, and delete the account. That promise cannot be kept without recording when you last signed in, so either the list was incomplete or the promise was empty. It was the list.
• How it was found. An independent review of the build plan, before any of it was built. Nothing had been implemented against the wrong list, which is the only reason this entry is short.
• What changed. The field is now in the list, in version 1.11, and it is named as an addition rather than slipped in. No account exists yet, so nothing was held under the old list and nothing needs deleting.

The uncomfortable part is what it says about “in full”. A list that claims completeness is a stronger promise than a list that does not, and this one was written before there was any code to check it against. The lesson taken is not to write weaker lists — it is that a completeness claim should be re-derived from the promises elsewhere on the page, every time one of them changes.

19 September 2026 — the privacy notice said there was no analytics, while an analytics script was on every page

This is the clearest case yet of the thing this page exists for: a promise that stopped being true without anyone editing the promise.

• What it said. The privacy notice read “No analytics. There is no analytics product on this site, first-party or otherwise” and “No third-party scripts … Every file this site loads is served from this domain.” The home page carried the short version: “no cookies, no analytics, no third-party scripts”.
• What was true. Earlier the same day, Cloudflare Web Analytics was switched on for this domain. It adds a small script to every page as the page leaves the network. Nine pages were checked and all nine carried it — including the seven tests, the Fingerprint, and the privacy notice itself.
• How it was found. Not by reading the page. Our deploy script compares what we publish against what is actually served, and one page came back 367 bytes larger than the file we had sent. The difference was the script.
• What it collects, checked against Cloudflare's own documentation rather than assumed: page views and where a visitor arrived from. It sets no cookie and uses no browser storage, it does not identify anyone by IP address or browser, and the IP address is discarded at the edge rather than stored. It has never had access to a single answer or score, because those never leave your browser.
• What changed. The privacy notice is now version 1.9 and says plainly that there is one analytics script and what it does. We did not remove the analytics — we decided we want it, and the honest response to wanting it is to say so rather than to keep a sentence that reads better.
• What is still wrong, today, as we publish this. We have decided the script belongs on the pages of text and not on the tests or on your Fingerprint. It is still on those pages while this entry goes up. That is one setting and it is being changed; this sentence stays here until it has been, and this entry will say so when it is done.

The uncomfortable part is the eleven hours. The script went on in the morning and this entry is written the same evening, which is fast — but the only reason it was hours rather than months is that a script compares served bytes against published bytes. Before today it compared status codes, which would have gone on saying 200 forever.

7 September 2026 (fifth entry) — we thanked people for an answer we never received

This is the worst kind of defect this site can have, because it made the page say something untrue to the person reading it.

• The two optional questions at the end of the word meaning result did not work at all. Pressing a button sent the answer to us, our own server rejected it, and the page displayed "Recorded. Thank you — that is genuinely useful." anyway.
• Why: the tally records which test an answer came from, and the list of permitted test names is enforced on our side. Word meaning was published without being added to that list, so every press returned an error. The page never checked.
• How long: from the moment word meaning was published on 7 September until this correction, the same day.
• What was lost: nothing of yours. No answer was stored, which is exactly the problem — you were told one had been.

Two things changed, and only one of them is the obvious fix. The obvious one is adding word meaning to the permitted list, and that is pending a separate deployment. The one that matters more is that the page no longer thanks you unless our server actually accepted it — if it does not, the page now says so and links here. A confirmation that fires whether or not the thing happened is not a confirmation, it is decoration, and it would have hidden this defect indefinitely.

Found by testing the endpoint directly while building the next instrument, not by using the page — because using the page shows you the thank-you either way. That is the lesson: a success message written by the client is not evidence that anything succeeded.

7 September 2026 (fourth entry) — and then we looked, and found three more

Writing the entry above prompted a sweep for the same defect elsewhere. It was live in three more places, and the worst was on the page that exists to prevent it.

• The instrument register said "There are two tests" and "It lists two tests" — while listing five. The register is the single source for how many tests this site has, and its prose disagreed with its own table, twice.
• The homepage said "four tests, four domains", when five tests and seven domains exist — and its list of tests was missing the fifth entirely, so a reader could not reach Word Meaning from the front page.
• The methodology page said "One test is now published", which was true on 3 September.

All four are fixed the same way: by not stating a count in prose at all. The register's table is now the only place the number lives, because a number written into a sentence goes stale without anyone editing it — the same reasoning that removed a live count from the privacy notice in version 1.4.

And it is now a check. Four occurrences of one defect class is not bad luck; it is a process choosing to keep a defect. The checker now fails if any page except this one states an inventory in prose. It found two false positives on its first run — this page's own quoted wording, and a structural phrase about one test per domain — and both were fixed in the check rather than in the pages, because both pages were right.

7 September 2026 (third entry) — the privacy notice named one test again

Third time for this exact sentence, and that is the interesting part. Under What we collect, the notice said "The digit span test runs entirely in your browser" — naming one instrument when five exist.

• Version 1.4 corrected a live count in this notice because a notice carrying a running number goes wrong on its own. Version 1.6 corrected the scope line for the same reason. This is the same defect, one paragraph further down, surviving both passes.
• It was never a wrong claim about what we collect — nothing is stored, and that was true throughout. It was an example that stopped describing the site around it.
• Fixed by removing the count rather than updating it: the sentence now reads "All five tests run entirely in your browser" and will be wrong again the day a sixth ships — so the real fix is that the register is the only place a number of tests is stated.

Found while publishing the notice's forward-looking section, not by auditing the notice. Which is the pattern worth naming: this page keeps recording a class of defect that a checker could catch, and the checker still does not catch it.

7 September 2026 (second entry) — the page explaining the arithmetic got the arithmetic wrong

Hours after publishing the number, the page that explains how it is calculated published a worked example that does not calculate.

• It said: seven digits, eleven grids, thirty matches and seven items unfolded give arms of 0.50, 0.56, 0.50 and 0.48 — "a number of 49".
• Their mean is 0.5104. The number is 51.
• The code was right the whole time. Type that example into the Fingerprint and it returns 51, which is what it always returned. The wrong number was in the prose explaining the code, not in the code — the one place a reader checking our working would have been misled by us while the product behaved correctly.

How it was found matters more than the slip. Not by re-reading the sentence, which had already been read several times, but by running the published example against the live page and comparing. A worked example is a test case, and it should be executed rather than proofread.

7 September 2026 — we said there would be no single number, and now there is one

This is the largest reversal on this page, and the founder asked for it explicitly rather than it being quietly slipped in. That distinction is written into how these decisions are supposed to work, so it is written here too.

• On 6 September we published a page titled "Why there is no single BrainScore yet", arguing that four scores in four different units could not be combined without a reference distribution we do not have. On 7 September a number appeared.
• What was wrong was not the argument. It was the scope of the question. That page was reasoning about a norm-referenced composite — a number that tells you where you sit among people. That still needs a reference sample, we still have none, and the four conditions that page set out are unchanged and still govern it.
• What it missed is that a number can be scaled against a task instead of against people. Each test has a floor — what guessing produces — and a ceiling the task was built with. Your position between them needs nobody else's results. That number now exists, and it is not a percentile.

One constant in it is a decision rather than a measurement, and it is marked as one on every page that uses it: the processing speed task is a rate with no natural ceiling, so we declared sixty correct — one match every 1.5 seconds for the full ninety seconds — as the top of that arm. It is not derived from anyone's performance.

Why we are telling you this rather than shipping the number quietly: the decision that said "no number" was recorded on 2 September with an explicit condition attached — that it could be revisited only through an open, evidence-based decision, never smuggled in. The full arithmetic, the weighting choice, and what the number still cannot say are published here.

6 September 2026 (fourth entry) — the privacy notice named one test while four exist

Found by the same method as the entry below: reading the live key store against the notice. This time it was the notice's own scope line.

• The notice said it covered "this website, including the digit span test". Written when there was one test. There are now four, and all four are covered — the sentence was narrower than the thing it described, which on a privacy notice is the wrong direction to be wrong in. It now says all four.
• A second counter left by an earlier version of the questions, b2:u:partly, was not named. Like tally:yes it holds one increment from our own testing. Both are now named in the notice with their values.

In the same change, and disclosed rather than slipped in: the two optional questions now appear after every test rather than after one, and the tally records which test an answer came from. That raises the maximum from seven numbers to seven per test. Nothing else about the request changed — still two short words from a closed list, with no result, no identifier and no time. The notice is at version 1.5.

6 September 2026 (later still) — two things we hold that this site did not describe

Found by reading the live key store against the privacy notice, which is not a thing we had done before. It is now written down as a document, so it can be done again by someone who was not here.

• The privacy notice ended a paragraph with "Today that means we hold one." We held two. The sentence was true when it was written on 5 September and became false the next time anybody pressed a button — which is the problem with putting a live count in a notice: it goes wrong on its own, with nobody editing it. The count is gone rather than corrected. The rule it illustrated — at most seven, and a counter does not exist until its button is pressed — is unchanged and still stated.
• And there is a third number we hold that the notice did not account for. A counter left over from an earlier version of those questions, holding a single increment from our own testing. Nothing on the site can add to it and nothing reads it. The notice said we hold "one for each button that can be pressed", and no button produces this one — so the notice was describing a store slightly smaller than the store. It is now named in the notice in full, including its name and its value. Removing it is a deliberate change to live data and will be made and recorded, not made quietly.

Neither is a change to what the site collects. What changed is that the page now accounts for everything in the store rather than for everything we meant to put there. The notice is at version 1.4.

6 September 2026 (later) — the symbol match page described its own scoring rule wrongly

Found while writing the page about why this test disagrees with other tests — which required re-reading our own procedure closely enough to describe it, and that is when the sentence stopped making sense.

• The result section said the within-run comparison was "ten seconds either side of the midpoint". It never was. The window is split at 45 seconds into two equal halves. The scoring rule further down the page said 45 seconds, the code splits at 45 seconds, and the caption the page prints under your result says 45 seconds — so one sentence disagreed with the other three, and the one that disagreed sat in the result section rather than in the scoring appendix. Nothing about the test changed and the instrument stays at v1.0: the number you got yesterday is the number you get today, and it was always computed from two 45-second halves. What was wrong was the description.
• And the working memory page still said "the question most people arrive with". Our voice document bans that phrase as rhetorical filler as well as an unmeasured claim about people — it was stripped from two pages on the day they were published and left standing on one written earlier. Same class, missed on the older surface: the third time today a fix has landed on one page of a pair. It now says the question the search results are full of, which is something we have actually looked at.
• The note shown when your second half is lower called that "the less common direction". We have measured nobody. We have no basis for saying how common either direction is, and saying so anyway is the exact failure this site was built to avoid. The frequency claim is gone; the mechanism — familiarity with the key pushes the other way — stays, along with an explicit statement that we do not know how often it happens.

6 September 2026 — three corrections from reading our own pages against each other

None of these was reported to us and none was found by looking at the page in question. All three surfaced from comparing one of our surfaces against another — which is the only method that has ever worked here, and is worth saying plainly.

• The digit span band labels named a kind of person. They read Memory Architect, Deep Coder, Fluent, Retentive and Building. Our own claims standard permits an evocative label only when it is attached to a task result, never as a standing description of the person — and those five failed that test, on a page that elsewhere tells you the number is not a verdict on you. They now describe what happened on the task. Found by comparing the digit span bands with the odd-one-out bands, which had been written later and correctly. No boundary moved and the instrument stayed at v1.2.
• Then the fix was itself incomplete, for several hours. The result copy was changed and the band table further down the same page was not, so the page carried a paragraph saying the labels had been replaced directly above a table still listing the old ones. A page that contradicts itself about its own scoring is worse than one that never claimed to have fixed anything. Corrected the same day; the two places now agree.
• The homepage said there was one test, four days after there were three. It also said one test, one subject area, and its list of what is on the site named one domain out of the three published. It had simply been left behind while pages were being built. The counts and the list now come from the register, which is the source for both. This is the entry surface, which makes it the worst page on the site to be wrong about how much of the site exists.

5 September 2026 — the test was easier than we said, in one specific case

This one is both a correction and a version change, which the policy above allows for and this is the first instance of.

• The digit span page publishes its timing: each digit on screen for 1000ms, with a 200ms gap. That was not true if you switched tabs. Browsers deliberately slow down timers in background tabs to save power. We measured our own live page with the tab hidden: each digit stayed up for about 2000ms instead of 1000ms, and the gap stretched from 200ms to about 1000ms. A longer look at each digit is an easier task — so anyone who switched away mid-sequence was taking a different, gentler test from the one the page describes, and nothing told them so. Fixed by adding a third exclusion rule: an attempt where the tab is hidden while the sequence is showing is now void, and that length is presented again. Because that changes an exclusion rule it is a version change, so digit span is now v1.2 and the methodology version is 1.2. We cannot tell which past results were affected, and there is nothing to re-score: we hold no results at all, which is the one place our reference sample of zero is an advantage.

How we found it is worth stating, because it was luck rather than diligence. We were driving the test through an automated browser to check that a new part of the result page rendered. The automation ran in a background tab, the test crawled, and the reason it crawled turned out to be a defect in the instrument rather than in the automation. A test nobody had taken in a hidden tab would not have surfaced this.

5 September 2026 — a claims sweep of every page

We checked all 57 factual assertions across all seven pages then published, against their sources. Most held. These did not.

• The working memory page said working memory and intelligence "are related in the research literature". We had not verified that, and it is the kind of claim we hold ourselves to checking. The sentence now says only that they are not the same thing and neither substitutes for the other, which is what we can stand behind.
• The digit span page said "the whole plausible range of adult performance spans about three digits" — true, but it was our own summary rather than something we had sourced. It now states the finding it came from: in the study that page cites, 80% of its 763 adults scored between 5 and 7 digits.
• The same page said the size of the memory limit "is still argued over in the literature". We had not checked the current state of that debate. It now points at what our own cited sources disagree about, which is a claim we can support.
• The privacy notice said we hold seven numbers. We hold at most seven — a counter does not exist until somebody has pressed that button. Today that means we hold one. Corrected to version 1.3, because a privacy notice that overstates what it holds is still a privacy notice that is wrong.

5 September 2026 — the methodology page, re-read against the site

• It said the test "stores nothing". It stores no results, which is what was meant and what the rest of that paragraph says — but the test does send one thing: your answer to the two optional questions at the end of the result, as a counter with nothing attached to it. The privacy notice has described that accurately since the test shipped; the methodology page did not mention it. Now corrected to "stores no results", with the optional questions described alongside. Nothing about the system changed — only that page's account of it.
• It claimed "almost every cognitive test online will tell you your percentile". We have not measured that, and we hold ourselves to not asserting things about the world we have not checked. The sentence has been rewritten to state what a percentile requires, which is the part we can stand behind. The point it was making survives; the unverified claim does not.

Neither of these changes a scoring rule, an exclusion rule, the domain model or the norming rung, so the methodology version stayed at 1.1.

5 September 2026 — a claim that was not made in words

• Our site icon was three bars rising left to right. Three ascending bars is the ordinary visual shorthand for improvement over time, and improvement is the one thing this site will not claim — it is the claim the FTC fined Lumosity over, and we hold that visual grammar makes claims as surely as sentences do. The bars are now uneven rather than ascending, which reads as three different measurements instead of a trend. This one is a judgement about how a picture reads, not a factual error, and we are listing it as a judgement. We think we were making a claim by accident.

What has not been corrected

No scoring rule, exclusion rule, band boundary, domain definition or norming rung has changed since publication. Every correction so far has been us describing our own work inaccurately, or asserting something about the world we had not checked — not the measurement itself being wrong.

That is the direction we expected, and it is worth stating plainly rather than leaving you to infer it from the list.